Hermes Agent X Roundup - August 7, 2026: Slash Command Expansion
NOTE: The X API v2 free tier and OpenRouter both returned HTTP 402 Payment Required during today's roundup run. All sourcing is from the GitHub API and release notes. No tweet embeds or AI-generated OG image this round - the featured image is a stdlib PNG placeholder.
The 48-hour window after v0.20.0's release brought four new slash commands and two STT hardening patches. The Herald Release landed August 3 with roughly 3,650 commits, 1,400 merged PRs, and 650-plus contributors across 1,200 closed issues. The follow-on work arrived in a tight cluster on August 6-7.
/learn - knowledge-base skills for large corpora
PR #80804 by teknium1, merged August 7 (+136/-8, 3 files).
The /learn command authors expansive knowledge-base skills for large prose sources - books, paper stacks, specifications, and documentation corpora. It produces a lean SKILL.md index plus per-chapter references/ files loaded on demand, instead of forcing everything into a single roughly 200-line file.
The pattern was ported from virgiliojr94/book-to-skill (MIT license, 17.7k stars). The implementation modifies the agent learn prompt (77 lines added) and adds 40 lines of test coverage. User-facing documentation was added to the skills page.
/refine - on-demand memory and skill review
PR #79682 by teknium1, merged August 6 (+175/-3, 9 files).
/refine [focus] lets users trigger the background memory and skill self-improvement review on demand. It uses the same fork the automatic post-turn loop uses, but is user-triggered and steerable with optional focus instructions. The concept was adapted from Prime Intellect's Prime-Agent continual harness, where Hermes' equivalent durable harness state is the memory and skill stores.
The change touches nine files: agent background review, the gateway and CLI layers, slash command routing, and 56 lines of test coverage for focus steering.
/heartbeat - idle session re-entry
PR #79681 by teknium1, merged August 6 (+927/-2, 11 files).
The largest PR in this cluster: /heartbeat every <interval> <prompt> gives a session a recurring instruction that re-enters the conversation as a normal user turn whenever the session is idle and the interval has elapsed. A new hermes_cli/heartbeat.py module at 332 lines handles the scheduling, with 187 lines of tests.
The concept came from Prime Intellect's Prime-Agent. It closes issue #62648 and is part four of tracker #79686. The gateway layer required 97 new lines to integrate the heartbeat into the run loop, and a new user guide page documents the feature.
Quality gates for /goal
PR #79680 by teknium1, merged August 6 (+656/-5, 7 files).
/goal gate add <command> attaches deterministic quality gates to the goal loop. When set, each command must exit 0 before the judge may declare a /goal done. The implementation adds 327 lines to hermes_cli/goals.py and 224 lines of test coverage in test_goal_gates.py.
This is part two of tracker #79686, closing #34196. It was also adapted from Prime Intellect's Prime-Agent autonomous-gate concept.
STT hardening: idle unload and silence trim
Two STT PRs by kshitijk4poor, both merged August 7:
PR #81027 (+411/-9, 5 files) adds idle unload for the local whisper model. The faster-whisper model singleton was previously loaded on the first voice message and never released - the base model holds roughly 370 MB and competes with local LLMs for GPU memory on multi-purpose machines. The fix unloads the model after a configurable idle window, with 283 lines of tests.
PR #77581 (+558/-7, 5 files) adds pre-upload silence trim for cloud STT providers (groq, openai, mistral, xai, elevenlabs, deepinfra). Local faster-whisper got Silero VAD in #73934, so silence never reaches the local model. Cloud providers had no equivalent - raw audio uploaded untouched. The new stt.cloud_trim config block strips leading and trailing silence with configurable thresholds, saving bandwidth and API cost for rambling voice notes.
What this cluster means
The four slash commands in this window extend Hermes from a stateless tool-calling agent toward a durable runtime that maintains memory, learns from large corpora, monitors its own idle state, and enforces deterministic quality gates. The STT hardening makes voice interaction viable for long-running sessions without wasting GPU memory or burning API credits on silence.
[^1]: teknium1. "feat(/learn): expansive knowledge-base skills for books and large corpora." GitHub PR #80804. August 7, 2026. [^2]: teknium1. "feat: /refine - run the memory/skill self-improvement review on demand." GitHub PR #79682. August 6, 2026. [^3]: teknium1. "feat: /heartbeat - recurring session re-entry prompt fired when idle." GitHub PR #79681. August 6, 2026. [^4]: teknium1. "feat(goals): quality gates - deterministic commands that must pass before /goal completes." GitHub PR #79680. August 6, 2026. [^5]: kshitijk4poor. "feat(stt): idle unload for local whisper model." GitHub PR #81027. August 7, 2026. [^6]: kshitijk4poor. "feat(stt): pre-upload silence trim for cloud providers." GitHub PR #77581. August 7, 2026.